Magic - Enterprise-Grade Open-Source AI Agent Platform
The first open-source all-in-one AI productivity platform for enterprises, combining generalist agents, workflow engine, IM, and online collaboration with security and control.
Evidence: README claims sandbox isolation, approval workflows, budget controls, but no implementation details or code evidence. Deductions: claims lack code support, no dependency security audit or vulnerability handling evidence.
Evidence: README description aligns with some code (e.g., test fixtures), but no error handling or failure message documentation. Deductions: dependency availability unverified, failure messages not mentioned.
Evidence: README clearly identifies target users (individuals, SMBs, large enterprises) and scenarios (e.g., cross-border e-commerce, customer service), but capability boundaries are vague. Deductions: trigger precision (e.g., approval conditions) not specified, environment fit (e.g., Windows support) incomplete.
Evidence: README, LICENSE, deployment guide link present, but missing CHANGELOG, version numbers, known limitations. Deductions: information architecture incomplete, install notes rely on external script, naming stability unverified.
Evidence: README claims output as ready-to-use artifacts (PPTs, dashboards, etc.), but no actual output examples or user feedback. Deductions: marginal value not quantified, cost-benefit analysis missing.
Evidence: Some README claims traceable to code (e.g., test fixtures), but lack independent verification. Deductions: cross-source corroboration insufficient, facts vs. inferences not clearly separated.
- Security claims (sandbox, approval, budget control) in README lack code evidence; further implementation review needed.
- License is a modified Apache 2.0 with commercial restrictions; assess compliance carefully.
- Install script (curl | bash) poses supply chain risk; review script contents.
What does this agent do, and when should you use it?
Magic is an enterprise-grade open-source AI Agent platform designed to overcome the limitations of personal AI tools in large organizations. It offers a unified solution for knowledge consolidation, security isolation, budget control, human-in-the-loop approval, and deliverable generation. The platform includes both a personal AI assistant and specialized Expert Agents, encapsulating internal systems (ERP/CRM/databases) into accessible digital employees. It supports self-hosted deployment via Docker and a one-line script, as well as a cloud service. Magic is fully compatible with the Anthropic Skills and OpenClaw Skills ecosystems. Agents run in sandboxed containers with VPC isolation and Sidecar proxies, ensuring data security. Output goes beyond text to include PPTs, dashboards, reports, and Excel files.
Magic performs automated AI-driven business operations: it deploys via curl -fsSL https://getmagicrew.sh | bash, which creates a cluster and starts services (UI at localhost:38080). Each Agent runs in a sandbox container with VPC isolation, communicating via private endpoints. The Sidecar proxy manages network traffic per user, enforcing resource isolation. High-risk operations trigger an approval workflow. The platform enforces per-department, per-user, and per-Agent budget caps. It integrates with internal systems through MCP, and its rendering framework converts AI outputs into polished artifacts. It supports IM integration (WeCom, DingTalk, Lark) for reporting progress.
- A cross-border e-commerce company operates with 8 people, each using personal assistants that invoke Expert Agents for product selection, advertising, customer service, and logistics, automating end-to-end processes, achieving the output of an 80-person team.
- A senior executive asks 'Can East China gross margin hold this quarter?' at 2 AM; the personal assistant connects to ERP and finance systems, invokes a finance Expert Agent, and delivers a live dashboard in 30 seconds.
- A customer service Expert Agent automatically responds to an urgent email from a New York client at 3 AM, understanding product details and return policies, providing a resolved ticket summary by morning.
- The legal Expert Agent reviews outbound contracts, flags risky clauses, and suggests revisions before clients see them, triggering human approval when necessary.
- A new hire accesses project management expertise via personal assistant on day one, leveraging industry knowledge and historical cases to reach full productivity within a week.
- A three-person go-global team uses a market expansion Expert Agent to research regulations, generate compliant product descriptions, manage listings on local platforms, and handle after-sales across 10 countries.
What are this agent's strengths and limitations?
- Strong enterprise security: sandbox containers, VPC isolation, Sidecar proxy, and multi-tenant data isolation.
- Human-in-the-loop approval for high-risk actions.
- Deliverables not just text: built-in rendering to PPT, dashboards, Excel, etc.
- Broad compatibility with Anthropic Skills and OpenClaw Skills ecosystems, easing migration.
- Requires Docker, and Windows support is not yet available, limiting some corporates.
- Platform-specific: deeply tied to Anthropic/OpenAI APIs; moving to other providers may require adaptation.
- Enterprise edition and advanced features may require a commercial license (cost).
- Self-hosting demands DevOps skills and external network access.
How do you install or deploy this agent?
Requirements: Docker and curl, on macOS or Linux. Run: curl -fsSL https://getmagicrew.sh | bash. The script handles cluster creation, infrastructure, and service deployment. After completion, access the Web UI at http://localhost:38080. To teardown, use magicrew teardown. See docs/en/development/deploy/docker.md for details.
How do you use this agent?
After installation, open the Web UI at http://localhost:38080. Create projects, add Agents, and configure budgets and approval policies. The personal AI assistant integrates with IM (WeCom, DingTalk, Lark), calendars, email, and internal systems. Invoke Expert Agents through natural language instructions. High-risk actions require human confirmation. For cloud version, sign up at https://www.letsmagic.cn (China) or https://www.magicrew.ai (international).
How does this agent compare with similar options?
Compared to OpenClaw (personal AI assistant), Magic adds enterprise controls: unified data hub, budget caps, approval flows, sandbox isolation, and deliverable generation, making it suitable for organizational deployment.